Optimal Prediction Model of Default Probability Based on Multiple Machine Learning Methods

Abstract

The prediction of the probability of default can help banks and other financial institutions to effectively identify and assess the potential default risk associated with family farms, thereby reducing losses due to bad debts. Although many methods are available for constructing models for the probability of default, the choice of optimal models is still inconclusive. Taking the survey data of 722 family farms in China Inner Mongolia as the empirical objects, 4 machine learning methods, including binary classification logistic regression, decision tree CART algorithm, random forest, and kernel support vector machine, were used to construct the default probability prediction model for family farms. By comparing and analyzing the four models, we found a better default probability prediction model to help financial institutions better audit the qualifications of family farms and reduce borrowing risks. The results showed that (1) the three models except logistic regression had strong prediction ability, which was higher than 90%, and the classification effect was good; and (2) the random forest model had the best prediction effect, the decision tree was the second, and the logistic regression was the worst.